AI Research

DeepMind opens live data behind its weather AI

Google DeepMind released SatWeave and PARDIG, live and historical data behind WeatherNext 3.

OpenAI Built a Kill Switch for Rogue Models

OpenAI launched Control Room to track research jobs and halt output within 35 seconds.

Databricks says context beats bigger models

Databricks says Genie with Ontology hit 91% accuracy with less than half the tokens by reading notebooks, not just tables.

Lazarus turns dormant neurons into backdoors

Microsoft Research says fewer than 10 dormant neurons can backdoor an AI agent until a hidden trigger fires.

AlphaGenome 2 doubles variant recall in cells

DeepMind says AlphaGenome 2 finds 30% of causal variants vs 16% before, with cell-type specific prediction.

Runway turns video brain into world action model

Praxis-1 pairs video prediction with action prediction to control real arms, topping RoboLab-120 among open models.

Modal says agent RL is ready for production

Modal detailed async agent RL with 1,000-fold sync savings and 15 percent cheaper rollouts.

Cline Engineer Shows Agents That Fix Themselves

Cline's Ara Khan demoed a harness that rewrites itself after a 77.5% vs 85% gap, plus a 17-hour run from 27%.

Google's Carbon hunts 0-days in two hours

DeepMind released Gemini 4 Carbon with top cyber scores, same pricing as Argon but deeper reasoning for defenders and builders.

Claude shipped boards that worked first try

Claude Fable 5.2 designed three working electronics prototypes from prompts, with no circuit fixes needed after assembly.

Google robots hit 96% on factory kitting

DeepMind says Gemini Robotics 2.5 learns dexterous tasks faster and hit 96% on Mercedes-Benz kitting with Apptronik Apollo 2.

DeepMind proved Signal code with 6M lines

Google DeepMind's AlphaProof Code generated 6M lines of Lean to verify Signal's crypto code in four weeks, with humans reviewing 1%.

Labs no longer buy labels, they buy worlds

Handshake AI told Latent Space labs now buy simulated workplaces, not labels, to train and test agents.

AI Research This Week: Cheaper Inference, Reliable Agents, Stale Instructions

New work advances low cost inference and gradients, audits early stopping, and probes agent rule compliance and stale instructions.

Agents start beating Wall Street on earnings

Samaya AI tested agents on 450 earnings calls and says GPT-6 Astra beat bias-corrected consensus on revenue.

Your Coding Agent Can't Undo Its Mistakes

Sah showed 84% task success but only 47% recovery, and EvoUndo rescued 180 of 197 broken self-improvements.

Cohere finds cheap planners beat big agent teams

Cohere's Nouha Dziri said planning-only teams matched full teams at half the cost, with cheap workers cutting bills 60%.

Anthropic says agents should dream offline

Anthropic says offline dreaming lifted Claude from 13% to 48%, and small, tested memories beat large logs.

OpenAI grades AI on real errands, not exams

OpenAI introduced LifeVal to grade AI on daily tasks like meals, budgets and childcare, not just coding.

Messy bench data is the real AI test

OpenAI and Ligo grade agents on raw bench data. Astra jumps to 66.7% but still misreads controls.

Netflix says subjective AI evals need two judges

Netflix's Cameron Wolfe tells AI Engineer that subjective AI tasks drown in noise and need human evals, tuned LM judges, and production AB tests.

Verification now proves agent Rust is correct for every input

Microsoft Research details PDD Kit proving agent-written Rust login code for every input instead of sampling tests.

Ricursive's 65B Ripple tops US open models on agentic coding

Ricursive says Ripple-65B Preview scores 80.52 on TerminalBench 2.1, beating larger U.S. open-weight models after a $32M scratch build.

LLMs Can Now Read and Rewrite Weights Directly

MIT CSAIL shows GPT-6 Astra reading and editing neural weights as text, fixing Qwen3.5-0.8B math from 7.9% to 48.1% by changing 0.016% of weights.

AI Research This Week: Autoregressive To Diffusion, Mixture-of-Experts Drafting

Research this week explores converting autoregressive models to diffusion language models, speculative decoding for mixture-of-experts, and how agent context pipelines lose information.

Pretraining Poison Is Now a Remote Code Bug

Pretrained Qwen3 agents hit 92.7% remote code execution from 0.1% pretraining poison that survived clean SFT, DPO and GRPO.

Genie 4 adds audio and live Gemini control

Google for Developers debuted Genie 4 with native audio, photorealistic visuals from Gemini Omni Flash and live Gemini steering for real-time worlds.

The Cheaper Model That Billed More

Vals AI's Rayan Krishnan told AI Engineer a cheaper list price can mean higher cost per task when weaker models burn extra tokens.

A Pico Can Dream Games, But Not Much Else

SheepSeb puts a looped world model and Stream AC actor on a Pico 2W, but exact-frame and dreamed-play checks show Pac-Man and Doom-lite still break.

Mathematicians argue whether AI solved math or just spoiled it

Lex Fridman gathered three mathematicians after AI claims on Navier-Stokes and quasi-Riemann, and the panel split on whether math is solved, spoiled, or just beginning.

Google's Gemini 4 Carbon pushes cyber defense first

Google DeepMind says Gemini 4 Carbon leads cyber defense benchmarks and will roll out first to defenders via its Fairwind Program.

OpenAI says AI proved three hard conjectures false

OpenAI claims AI-generated, Lean-checked proofs for local smoothing, ETH and QCD, with three different public release models.

OpenAI cuts price and pushes speed with Terra

OpenAI introduced GPT-6.1 Terra as a faster, cheaper model for visual reasoning and computer use, priced 50% below its predecessor.

AI Research This Week: Long Context Scaling, Synthetic Data Laws, Tool Fault Fragility

This week research tackles cheaper scaling for long context and synthetic data and mixture of experts, and fragility in tool using agents where scale does not guarantee reliability.

We know less about AI than we did three years ago

CNN pressed AI builders on the black box problem, and one answer stood out: we know less now than in 2023 as models outpace our understanding.

AI Research This Week: Model-Specific Reasoning Bias, Generative Reward Scaling

New work finds model-specific bias in reasoning, scaling rules for generative reward models, and training-free pool sharing for sparse attention alongside multi-agent and tool market studies.

Frontier models still fail the buy-side memo test

Google DeepMind and LinqAlpha showed why frontier models fail as buy-side analysts and previewed Mandate Bench, a 50-task memo benchmark.

Tone Bench grades slop but only in one voice

Towards AI founder launches Tone Bench to measure slop in his own voice, but a single-voice test and biased judges leave general claims unproven.

MiniMax says faster is not always finished

MiniMax's Morgan Suo says M3.1 Flash is faster but urges testing the whole research job, not tokens per second, before switching models.

MiniMax Pushes Flash to 200 TPS at 70% Less Cost

MiniMax says M3.1 Flash, tested as Space Bunny, hits top agentic benchmarks at $0.15 input with 200 TPS output.

Gemini 4 Argon API Pricing Pressures Inference Economics at Two Dollars

Google opens Argon after gated defender trial

DeepMind moves Gemini 4 Argon from Fairwind defender trial to broad rollout, detailing layered misuse and misalignment defenses.

Mistral says 209M Timestral leads on three benchmarks

Mistral says its 209M Timestral family leads GIFT-Eval, TIME and fev-bench, with models from 5M to 209M that run on CPUs.

DeepSeek ships faster Flash with shorter outputs

DeepSeek's V4.1-Flash-1017 claims faster agentic execution with shorter outputs and drops the Jinja chat template for a new encoding toolkit.

Wisedocs says insurance claims still break AI

Wisedocs' Denys Linkov told AI Engineer that fragmented, fax-heavy claim files remain AI's hardest data problem, even as its MLCR benchmark jumps to 75%.

DeepSeek V4.1 Flash is cheap, fast and visual

DeepSeek V4.1 Flash 1017 brings native vision and CED architecture with 8B/16B active parameters and 1M context at $0.10/$0.50.

Meta bets Muse Wonder can catch OpenAI, Anthropic

Meta pitched Muse Wonder as frontier and on par with ChatGPT and Claude, but Bloomberg Tech guests see consumer strength and enterprise hurdles.

Small editors learn to rewrite agents at test time

Zhang et al. train a 4B proposer to edit frozen-agent harnesses from execution feedback, beating a 35B teacher on held-out Reasoning Gym families.

Weco's Aiden won Parameter Golf on 4% of compute

Weco AI detailed how its Aiden agent won OpenAI's 16MB, 10-minute parameter golf with seven merged records on 4% of compute.

NYU Mathematician Says OpenAI Copied His Work

NYU professor Tristan Buckmaster told CNBC Television OpenAI reproduced his Euler work after he revealed his name, casting doubt on its Navier-Stokes claim.